Misunderstanding the Machine: Munich's War on Generative AI

When the Munich Regional Court ruled against OpenAI in November 2025, finding that ChatGPT’s internal “memorisation” of German song lyrics constituted copyright infringement (our comment available HERE), we thought it spelled trouble for the already struggling European tech sector. But the court’s recent follow-up decision from July 2026, holding American AI music generator Suno Inc. directly liable for training on GEMA’s musical repertoire, confirms this troubling precedent and threatens to turn it into a real existential threat for European artificial intelligence. If higher appellate courts do not overturn this line of judicial reasoning, Europe risks turning its digital ecosystem into a technological desert, isolated from global innovation and governed by a legal framework completely detached from technical reality in order to protect outdated concepts and legacy interests.

At the heart of the Munich rulings lies a fundamental misunderstanding of how generative AI models are created and how they work. Judges have desperately attempted to stretch 20th-century statutory concepts of “mechanical reproduction” to fit 21st-century technology, that consists in statistical representations of complex relationships based on enormous datasets. It is a technical fact, not an opinion, that large language models and neural audio architectures do not store discrete, addressable files, digital copies, or hidden audio tracks anywhere within their weights, or even parts of them. Model weights are floating-point numbers – mathematical representations of abstract patterns, correlations, and probabilities across vast datasets. By ruling that a model’s ability to output a familiar melody or text snippet when prompted proves “memorisation” inside the model weights, the court has abused its discretion in order to construct a legal fiction of reproduction, where there is none. It thus treats statistical likelihood as physical duplication, ostensibly deligalising the very mathematical process that is the basis for AI. A pencil is capable of producing a copy of a copyrighted drawing, but no one in their right mind would find the pencil factory liable for such an act, for which the pencil is neither designed nor intended. 

This misapprehension effectively guts Europe’s Text and Data Mining (TDM) exceptions. Under Section 44b of the German Copyright Act, which implements Article 4 of the EU Copyright in the Digital Single Market Directive, commercial actors were explicitly granted the right to mine (ie. analyse) data for analytical purposes unless rightsholders expressly opted out. The Munich court effectively dismantled this important safeguard by splitting AI processing into artificial phases: claiming TDM covers only the initial scraping and format conversion, but not the actual training and retention of the model weights. If the mere retention of abstract patterns within model weights counts as an unauthorized reproduction, then the TDM exception is completely meaningless for generative model creators. No commercial entity can build a competitive foundational model if every statistical training run requires negotiating proactive, upfront licenses with every collecting society and rightsholder on the planet. Independent research has confirmed that models trained on smaller, purely licenced or public domain datasets, are less capable and cannot compete.

Potentially equally damaging is the court’s aggressive attempt to project European copyright law across foreign borders. Suno trained its models on infrastructure located entirely within the United States. Yet, the Munich judges asserted jurisdiction because the end product is accessible to users in Germany, flatly rejecting the argument that this constitutes “fair use” in the US. This extraterritorial overreach shreds the long-standing principle of territoriality in copyright law, equating model training with model use. In the current geopolitical situation, it is quite voluntarist to assume the European Union can unilaterally impose its copyright standards in relation to acts carried out in other jurisdictions. It is not likely that foreign AI developers will sign punitive licensing deals or open their financial books to European rights-management societies under threat of profit disgorgement. A far more likely outcome is that they will simply geofence their state-of-the-art models, depriving European users of access – a case of such denial recently sent shockwaves through global tech and policy circles. The current reality is that Europe has no comparable models to substitute foreign offerings with. Given this line of legal thinking prevails it will most likely never have them. 

For European model creators, this legal climate is catastrophic. While US and Asian developers operate under flexible legal regimes that allow rapid scaling, European AI startups risk being trapped under a mandatory “license-first” regime before writing a single line of code. The immediate result will not be a gold rush of licensing revenue for legacy rightsholders, but an accelerated flight of tech investment and development out of the bloc – which is already happening for years. By trying to protect existing legacy monopolies through an overstretched interpretation of copyright, European judges are effectively condemning the continent’s AI ambitions to permanent obsolescence.

While this is not a foregone outcome and there are many more legal hurdles along the way, which could mean years of litigation, the momentum and recent rumblings from Brussels, where the EC is considering a “targeted legislative instrument” to address the issues related to AI model training, give all European innovators serious causes for concern.

If you wish to help us avoid this future, consider joining the Copyright for AI Coalition. Contact directly Marcin Olender, Head of Copyright Advocacy marcin@aichamber.eu to learn more.